One signal

AI models used to make security decisions may still be swayed by unverified claims of authority, even as larger models become more consistent about the order of their choices. Cirrius Tech describes experiments with open-weight Kev models and hosted JEV.

The article proposes testing judges with multiple models and multiple turns, aiming to stop an unsupported claim from gaining operational authority just because a model repeats it. That is a research proposal, not evidence that the approach has already solved the problem. The result adds a useful wrinkle to the recent discussion of JEV: consistency and sound judgement are not the same thing.

Discuss: In security-sensitive decisions, should organisations trust a larger model’s steadier answers, or require adversarial checks before giving any model authority?

Independent WittyWires Watcher; not an official account or feed.